extern/CUT3R/config/dpt_512_vary_4_64.yaml

model: "ARCroco3DStereo(ARCroco3DStereoConfig(freeze='encoder', state_size=768, state_pe='2d', pos_embed='RoPE100', rgb_head=True, pose_head=True, patch_embed_cls='ManyAR_PatchEmbed', img_size=(512, 512), head_type='dpt', output_mode='pts3d+pose', depth_mode=('exp', -inf, inf), conf_mode=('exp', 1, inf), pose_mode=('exp', -inf, inf), enc_embed_dim=1024, enc_depth=24, enc_num_heads=16, dec_embed_dim=768, dec_depth=12, dec_num_heads=12, landscape_only=False))"
pretrained: cut3r_512_dpt_4_64.pth
load_only_encoder: False
long_context: True
fixed_length: False
resume: null
benchmark: False
num_views : 64
num_test_views : 4
n_corres_train: 0
n_corres_test: 0

train_criterion: ConfLoss(Regr3DPoseBatchList(L21, norm_mode='?avg_dis'), alpha=0.2) + RGBLoss(MSE)
test_criterion: Regr3DPose(L21, norm_mode='?avg_dis', gt_scale=True, sky_loss_value=0) + Regr3DPose_ScaleInv(L21, norm_mode='?avg_dis', gt_scale=True, sky_loss_value=0) + RGBLoss(L21)

resolution: [(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)]

allow_repeat: True
dataset1: Co3d_Multi(allow_repeat=${allow_repeat}, split='train', ROOT='../../data/dust3r_data/processed_co3d/', aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset2: WildRGBD_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/dust3r_data/processed_wildrgbd", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})

dataset3: ARKitScenes_Multi(allow_repeat=${allow_repeat}, split='train', ROOT='../../data/dust3r_data/processed_arkitscenes/', aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset4: ARKitScenesHighRes_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/dust3r_data/processed_arkitscenes_highres", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset5: ScanNetpp_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/dust3r_data/processed_scannetpp/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset6: ScanNet_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/dust3r_data/processed_scannet/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset7: HyperSim_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_hypersim", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})

dataset8: BlendedMVS_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/dust3r_data/processed_blendedmvs/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset9: MegaDepth_Multi(allow_repeat=${allow_repeat}, split="train", ROOT="../../data/dust3r_data/processed_megadepth", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset10: MapFree_Multi(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/mast3r_data/processed_mapfree/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset11: Waymo_Multi(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/dust3r_data/processed_waymo/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset12: VirtualKITTI2_Multi(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/mast3r_data/processed_vkitti", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset13: UnReal4K_Multi(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/mast3r_data/processed_unreal4k/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset14: TartanAir_Multi(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/mast3r_data/processed_tartanair/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})

dataset15: DL3DV_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_dl3dv", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})

dataset16: Cop3D_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_cop3d/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset17: MVImgNet_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_mvimgnet/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset18: RE10K_Multi(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/custom_data/processed_re10k/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset19: OmniObject3D_Multi(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/custom_data/processed_omniobject3d/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})

dataset20: ThreeDKenBurns(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/custom_data/processed_3dkb/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset21: IRS(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/custom_data/processed_irs/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset22: SynScapes(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/custom_data/processed_synscapes/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset23: UrbanSyn(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/custom_data/processed_urbansyn/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset24: EDEN_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_eden", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset25: SmartPortraits_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_smartportraits", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})

dataset26: DynamicReplica(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_dynamic_replica/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset27: Spring(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/custom_data/processed_spring/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset28: BEDLAM_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_bedlam", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset29: MVS_Synth_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_mvs_synth", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset30: PointOdyssey_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_point_odyssey", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})

dataset31: UASOL_Multi(allow_repeat=${allow_repeat}, split='train', ROOT="../../data/custom_data/processed_uasol", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})
dataset32: MP3D_Multi(allow_repeat=${allow_repeat}, split=None, ROOT="../../data/custom_data/processed_mp3d/", aug_crop=16, resolution=[(512, 384), (512, 336), (512, 288), (512, 256), (512, 208), (512, 144), (384, 512), (336, 512), (288, 512), (256, 512)], transform=SeqColorJitter, num_views=${num_views}, n_corres=${n_corres_train})

train_dataset: 44800 @ ${dataset1} + 56000 @ ${dataset2} + 56000 @ ${dataset3} + 22400 @ ${dataset4} 
  + 16800 @ ${dataset5} + 22400 @ ${dataset6} + 11200 @ ${dataset7} 
  + 22400 @ ${dataset8} + 22400 @ ${dataset9} + 84000 @ ${dataset10} + 56000 @ ${dataset11}
  + 5600 @ ${dataset12} + 168 @ ${dataset13} + 56000 @ ${dataset14} + 84000 @ ${dataset15}
  + 480 @ ${dataset16} + 19200 @ ${dataset17} + 4800 @ ${dataset18} + 38400 @ ${dataset19}
  + 26400 @ ${dataset26} + 1200 @ ${dataset27} + 36000 @ ${dataset28} + 2400 @ ${dataset29}
  + 24000 @ ${dataset30} + 14400 @ ${dataset31} + 28800 @ ${dataset32}
test_dataset: 1000 @ ARKitScenes_Multi(split='test', ROOT='../../data/dust3r_data/processed_arkitscenes/', resolution=(512, 384), num_views=${num_test_views}, seed=42, n_corres=${n_corres_test})

seed: 0
batch_size: 4
accum_iter: 4
gradient_checkpointing: True
epochs: 10
start_epoch: 0
weight_decay: 0.05
lr: 1e-6
min_lr: 1e-7
warmup_epochs: 0.5
amp: 1

num_workers: 4
world_size: 1
local-rank: -1
dist_url: 'env://'
rank: 0
gpu: 0
distributed: False
dist_backend: 'nccl'

eval_freq: 1
save_freq: 0.1
keep_freq: 1
print_freq: 10
print_img_freq: 50000000
num_imgs_vis: 4
save_dir: 'checkpoints'
exp_name: 'dpt_512_vary_4_64'
task: 'cut3r'
logdir: ./${save_dir}/${exp_name}/logs
output_dir: ./${save_dir}/${exp_name}/
hydra:
  verbose: True
  run:
    dir: ./${save_dir}/${exp_name}